perf tool/perf_event_open and uncore events

perf tool/perf_event_open and uncore events

[Manuel Selva]

Hi all,

I am planning to use the perf tool and maybe directly the 
perf_event_open system call from code to get access to uncore counters . 
I am working on a dual socket NUMA platform with 2 Intel Xeon X5650 cpus 
(Westmer-EP familly and 06_2CH cpu number). I started my experiments 
with the UNC_QMC_NORMAL_READS.ANY counter only. I want to measure memory 
traffic for each one of my 2 memory controllers.

In this context I got the following questions without being able to 
answer to them:

+ It seems impossible to have per thread mode with this uncore event. I 
have to specify the -a or the -C option to the perf tool to get some 
results. Looking at the perf_event_open system call happening behind the 
scene, and calling it myself from code, I noticed that the call 
succeeded only when the pid parameter is -1 and the cpu one is different 
from -1. I guess that per thread counting is not available for uncore 
events because it's impossible for the hardware to measure only the 
uncore events generated by a given thread (because by definition uncore 
is shared by all the cores associated to it) Am I right ?

+ How the cpu parameter to perf_event_open should be interpreted for 
uncore events ? On my platform I have 12 cores (my two cpus are hexa 
core ones and hyper threading is disabled). Does providing a cpu value 
of 5 to the system call along with an uncore event means counting the 
events on the uncore associated to the core 5 ?

Thanks for your inputs,

-- 
Manu

Re: perf tool/perf_event_open and uncore events

[Andi Kleen]

Manuel Selva <selva.manuel@gmail.com> writes:
>
> + It seems impossible to have per thread mode with this uncore
> event. I have to specify the -a or the -C option to the perf tool to
> get some results. Looking at the perf_event_open system call happening
> behind the scene, and calling it myself from code, I noticed that the
> call succeeded only when the pid parameter is -1 and the cpu one is
> different from -1. I guess that per thread counting is not available
> for uncore events because it's impossible for the hardware to measure
> only the uncore events generated by a given thread (because by
> definition uncore is shared by all the cores associated to it) Am I
> right ?

Yes uncore counting is per socket counting.
>
> + How the cpu parameter to perf_event_open should be interpreted for
> uncore events ? On my platform I have 12 cores (my two cpus are hexa
> core ones and hyper threading is disabled). Does providing a cpu value
> of 5 to the system call along with an uncore event means counting the
> events on the uncore associated to the core 5 ?

Yes, the socket associated with the CPU is used.

-Andi


-- 
ak@linux.intel.com -- Speaking for myself only